Training code for resnet34.a1_in1k

[vvhj3357816826-dotcom]

I attempted to continue training on the resnet34.a1_in1k weights using the following parameters: --data /dataset/imagenet --model resnet34rer -b 128 --lr 1e-6 --weight-decay 0.01 --aa rand-m7-mstd0.5-inc1 --cutmix 0.2 --color-jitter 0. --amp --warmup-lr 1e-6 --experiment resnet34rer --opt lamb --initial-checkpoint a1weight/resnet34_a1_0-46f8f793.pth --bce-loss --grad-accum-steps 2

However, it was found that the learning is not continuous, and the accuracy has a 0.5% decrease.

Could you please provide the relevant scripts for this training?

[rwightman]

@vvhj3357816826-dotcom I don't see the point here? There is no point to continue training of the finished model like that, it's expected that it will end up worse. There are other threads here you can find on the hparams for training from scratch. The paper used a global batch size of 2048 so 128 you need to adjust the learning rate and it could take some tuning to find the sweet spot for resulting LR / weight decay.